基于注意力机制和元特征二次重加权的小样本目标检测  

Few-shot object detection based on attention mechanism and secondary reweighting of meta-features

作　　者：林润超 黄荣 董爱华[1] LIN Runchao;HUANG Rong;DONG Aihua(College of Information Science and Technology,Donghua University,Shanghai 201620,China)

机构地区：[1]东华大学信息科学与技术学院,上海201620

出　　处：《计算机应用》2022年第10期3025-3032,共8页journal of Computer Applications

基　　金：国家重点研发计划项目(2019YFC1521300)。

摘　　要：在基于迁移学习的小样本目标检测任务中,由于缺乏关注图像中待检测目标的注意力机制,所以现有模型对于待检测目标周边背景区域的抑制能力不强,且在迁移学习过程中通常需要对元特征进行微调来实现跨域共享,这将引起元特征偏移,从而导致模型对大样本图像检测能力的下降。针对上述问题,基于注意力机制和元特征二次重加权机制,提出改进的元特征迁移模型Up-YOLOv3。首先,在原始元特征迁移模型Base-YOLOv2中引入基于卷积块注意力模块(CBAM)的注意力机制,使特征提取网络聚焦于图像中的目标区域并关注图像目标类别的细节特征,从而提升模型对小样本图像目标的检测性能;其次,引入基于压缩-激励(SE)的元特征二次重加权模块(SE-SMFR)对大样本图像的元特征进行二次重加权,以获取二次重加权元特征,使模型在提升小样本目标检测性能的同时也能减小大样本图像元特征信息的权重偏移。实验结果表明,在PASCAL VOC2007/2012数据集上,相较于Base-YOLOv2,Up-YOLOv3针对小样本图像检测的平均准确率均值(mAP)提升了2.3~9.1个百分点;相较于原始的基于YOLOv3元特征迁移模型Base-YOLOv3,Up-YOLOv3针对大样本图像的mAP提升了1.8~2.4个百分点。可见,改进后模型对不同类别的大样本图像和小样本图像均具有良好的泛化能力和鲁棒性。In the few-shot object detection task based on transfer learning, due to the lack of attention mechanism to focus on the object to be detected in the image, the ability of the existing models to suppress the surrounding background area of the object is not strong, and in the process of transfer learning, it is usually necessary to fine-tune the meta-features to achieve cross-domain sharing, which will cause meta-feature shift, and lead to the decline of the model’s ability to detect large-sample images. To solve the above problems, an improved meta-feature transfer model Up-YOLOv3 based on the attention mechanism and the meta-feature secondary reweighting mechanism was proposed. Firstly, the Convolution Block Attention Module(CBAM)-based attention mechanism was introduced in the original meta-feature transfer model Base-YOLOv2, so that the feature extraction network was able to focus on the object area in the image and pay attention to the detailed features of the image object class, thereby improving the model’s detection performance for few-shot image objects.Then, the Squeeze and Excitation-Secondary Meta-Feature Reweighting(SE-SMFR) module was introduced to reweight the meta-features of the large-sample image for the second time in order to obtain the secondary reweighted meta-features, so that the model was not only able to improve the performance of few-shot object detection, but also able to reduce the weight shift of the meta-feature information of the large-sample image. Experimental results on PASCAL VOC2007/2012 dataset show that, compared with Base-YOLOv2, Up-YOLOv3 has the detection mean Average Precision(mAP) for few-shot object images increased by 2. 3 to 9. 1 percentage points;compared with the original meta-feature transfer model based on YOLOv3 Base-YOLOv3, mAP for large-sample object images increased by 1. 8 to 2. 4 percentage points. It can be seen that the improved model has good generalization ability and robustness for both large-sample images and few-shot images of different classes.

关 键 词：小样本目标检测 元特征迁移 特征重加权 注意力机制 二次重加权 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]